|
Abstract:
|
When the cells of a contingency table are unplanned domains, small realized sample sizes can cause direct survey estimators to be unreliable. We develop a procedure that obtains more stable estimators of the cell totals and proportions in a two-way table under an assumption that the margins of the table are well estimated. The method preserves the direct estimators of the marginal totals and incorporates information from a previous census. The procedure was developed for the Canadian Labour Force Survey as a way to improve occupational detail at the province level. In a simulation, the predictor achieves a smaller mean squared error than the direct estimator. Due to variability in the direct estimators of the marginal totals, the reduction in the MSE is greater for proportions than for totals. Empirical coverages of nominal 95% prediction intervals are betweeen 93% and 96%.
|
